The Agent Skills Directory

Indirect Prompt Injection Surface: The skill ingests untrusted data from Azure resource tags and Advisor recommendations to generate reports. This represents a potential surface for indirect prompt injection; however, the skill mitigates risk by prohibiting automated changes and requiring explicit human confirmation for any suggested actions (such as resizing or deleting resources).
Ingestion points: resource inventory tags and Advisor recommendation properties in references/optimization.md and references/resource-graph.md.
Boundary markers: The instructions emphasize preserving recommendation text verbatim rather than interpreting it, though specific delimiters for external content are not mandated.
Capability inventory: The skill can create files (create_file) and execute shell commands (az cli, powershell) as fallbacks.
Sanitization: The instructions focus on validated server-side projection and aggregation rather than local parsing.
Dynamic Query Execution: The workflow utilizes dynamic KQL generation for Azure Resource Graph queries. To ensure query integrity, the skill includes a mandatory validation step (validate_query) before any query is executed (execute_query), which is a standard security practice for dynamic data retrieval.
Restricted Command Execution: The skill allows for the use of authenticated shell clients (Azure CLI or PowerShell) as a fallback mechanism when standard tool operations are unavailable. These are restricted to vendor-supported tools, and the instructions explicitly forbid using them for local data aggregation or exposing authentication tokens.

cost-optimization